forge 0.7.3 — native-first proxy

[antoinezambelli]

forge 0.7.3 is out — pip install -U forge-guardrails

The proxy work in 0.7.x drew more interest than we expected, and that surge stress-tested the prompt-injection path harder than it had been. Taking a beat to sit with it, that path wasn't holding to the standard I want — and alongside ongoing dogfooding against more complex agentic systems, and where modern local models now sit on native function calling, it pushed us to a native-first stance.

So 0.7.3 makes native the default and the optimized path: the proxy forwards your OpenAI tool/message payloads to the backend verbatim. Prompt-injection isn't gone — it's an explicit opt-in (--backend-capability prompt) for llama.cpp/llamafile backends without a function-calling template. It's the right tool when a model has no FC template, but under more complex agentic scenarios models tend to struggle with it, so it's a fallback now, not the front door.

⚠️ One breaking change

The proxy --mode {native,prompt} flag is renamed to --backend-capability {native,prompt} (it collided with the proxy's managed/external deployment mode, and the new name reflects that the choice is fixed at startup, not switched at runtime). No alias — --mode was only a week old.

• --mode native → drop it (native is the default) or --backend-capability native
• --mode prompt → --backend-capability prompt

Sorry for the churn so soon after shipping it — but ripping the band-aid off now, while things are still picking up, beats carrying a confusingly-named flag and a compatibility shim forever. Better in the long run.

Also in this release

• An OpenAICompatClient for any OpenAI-compatible endpoint (thanks @lucasgerads)
• A configurable proxy --backend-timeout
• A batch of proxy hardening and fixes (thanks @hobostay, @alexandergunnarson)

Full notes: https://github.com/antoinezambelli/forge/releases/tag/v0.7.3

As always, feedback welcome — especially from anyone running the proxy in front of a local model for agentic work.